基于深度信息与尺度空间的仿射不变特征检测*

介绍了一种利用深度信息的仿射区域检测器。这种方法在视角变换的情况下能自动检测出图景中同一物理区域，为后续的识别算法提供了坚实的特征检测基础，在计算机视觉领域有广阔的应用前景。该方法是基于尺度空间理论，这个理论已经在自动尺度选择中有较成熟的应用。提出了利用深度信息估计出3D物体模型的算法，并生成相应的仿射不变的高斯尺度空间，并给出从3D到2D的投射变换的高精度估计方法，以补偿投射变换造成的扭曲形变。因此对特征检测的可靠性将有明显的提高。为了评估本算法的鲁棒性，进行了不同视角的真实图片与合成图片的实验，并与其

Affine invariant feature detection based on scale space analysis and depth information

This article presented an affine region detector using the depth information of the image. Is was concerned with a method for automatically detecting distinctive image patches covering identical physical part of scene even when the change of viewpoint. It could provide robust feature detection for further recognition process and could be widely used in the field of computer vision. The method was based on scale-space theory, which was also used for automatic scale selection. It proposed an algorithm, which could estimate the 3D part of the object and generate corresponding affine invariant Gaussian scale space with the depth information. This improved the reliability of feature detection, since the perspective transformation from 3D to 2D could be relative precisely simulated, which could compensate for the distortions due to perspective transformation. In order to test the robustness of the algorithm, experiments on real and synthetic images with known orientation were taken. The comparison with other affine invariant detectors was demonstrated as result.